Recognition of a stream of random events

The paper is concerned with recognition of a flow of events which cannot be observed directly (of a finite number of classes) that occur at random times on the axis of discrete time. Every event is assumed to cause a change in the probabilistic properties of a sequence of the recorded values of the feature vector in comparison with past values. That change is related to the class of the subsequent event. For the proposed class of loss functions the recognition problem is reducible to an extremal problem of optimal decomposition of the realization fragment in the observation interval.